Computer Vision
Modern computer vision means convolutional neural networks (CNNs). They are built from a few repeating bricks: a convolutional layer extracts features with learnable filters, a pooling layer shrinks the feature map with no learning, and augmentation artificially enlarges the dataset. The classic architectures — Inception and ResNet — are assembled from the same bricks.
The topic looks simple but is full of traps interviewers love. A filter sees all input channels, not one. Pooling has no learnable parameters. Augmentation that changes the label (flipping the digit 6 into 9) corrupts the ground truth, and "fine-tuning on the test set" is a leak, not a trick. The full map lives in the layers below.
Topic map
- Convolutional layer — learnable filters sliding over the input and spanning all channels; how the filter count becomes the output-channel count.
- Pooling layer — max/average downsampling with no learnable parameters, shift invariance, and a growing receptive field.
- Data augmentation — enlarging the dataset with transforms during training and averaging over views at inference (TTA).
- CNN architectures — the Inception module with its 1×1 bottleneck and the ResNet residual block against vanishing gradients.
Common traps
| Mistake | Consequence |
|---|---|
| Thinking a filter sees one channel | A filter spans all input channels; otherwise the convolution mechanics are wrong |
| Confusing a conv layer with a fully-connected layer | You lose weight sharing and the sliding window — the whole point of convolution |
| Thinking pooling has learnable weights | Pooling applies a fixed function; there are no parameters |
| Thinking a 1×1 convolution changes the spatial size | It changes the channel count, not width and height |
Forgetting the +x term in a residual block | The gradient highway disappears — the deep network will not train |
| Fine-tuning on the test set as "augmentation" | Label leakage and an inflated, dishonest score |
Interview relevance
CNNs are asked to check whether you separate learnable mechanisms from fixed ones and understand why the network is built the way it is. A candidate who explains the residual block as "an identity skip F(x) + x giving gradients a direct path to early layers" immediately gets ahead of "well, there are some skip connections".
Typical checks:
- How a convolution works — the filter spans all channels, slides, shares weights; filter count = output-channel count.
- How pooling differs from convolution — a fixed function with no parameters, and why it is needed at all.
- How to squeeze out a score without retraining (TTA) and which augmentations are safe.
- How Inception and ResNet let networks go deep.
Common wrong answer: "pooling is a learnable layer that recovers detail". In fact pooling reduces the spatial size with a fixed function (max/average) and holds no learnable parameters at all.